Skip to content

Identifiability

Prime #
903
Origin domain
Statistics & Experimental Design
Subdomain
causal inference and system identification → Statistics & Experimental Design

Core Idea

Identifiability is the structural condition under which an internal unknown — a parameter, mechanism, causal effect, hidden state, or latent variable — is in principle recoverable from the observable signal the system makes available. The defining commitment is a uniqueness claim: across the space of admissible internal models, the mapping from internals to observations is one-to-one within whatever subspace the observation can see. When two distinct internal configurations produce the same observable distribution, the underlying object is unidentified, and no amount of additional data of the same kind can distinguish them — only a structural intervention (a new measurement channel, an experimental manipulation, a parametric restriction, a prior commitment) can.

The pattern has the same shape across substrates: there is a target object (the thing you want to know), an observation map (the channel the world makes available), and an equivalence class on the target space induced by the observation map. Identifiability is the property that this equivalence class is a singleton for the target value of interest. The diagnostic question — could two structurally distinct internals produce identical observations? — is substrate-neutral, and its negative answer ("yes, they could") tells you, before any data arrive, that the model is misspecified for the question being asked.

What identifiability provides as a prime is the separation of an information-theoretic upper bound from estimation difficulty. Failure to identify is not solved by more data, better estimators, or larger computation — those help only when the underlying object is identifiable in the first place. Estimation noise lives downstream; identifiability gates whether the destination exists at all. The signature is bare relational structure, recognized across fields without importing any home vocabulary or normative load, which is why it reads as fully structural even though its heaviest usage clusters in formal and statistical domains.

How would you explain it like I'm…

Can You Even Tell?

Imagine I tell you two numbers add up to 10, and ask which two numbers I started with. You can't know — it could be 4 and 6, or 1 and 9. The answer is hidden because adding hides which pieces went in. Identifiability is asking: can you actually figure out the hidden answer from what you're allowed to see, or is it impossible no matter how hard you look?

Is the Answer Reachable?

Sometimes you want to know something hidden inside a system, but all you get to see is what comes out. Identifiability asks whether the hidden thing can even be figured out from what you can observe. If two totally different hidden setups would produce the exact same observations, then you simply cannot tell them apart — and collecting more of the same kind of data won't help, because the data can never separate them. The only fix is to change what you can see: a new kind of measurement, a real experiment, or an extra assumption. So identifiability is about whether the answer is reachable at all, before you even worry about getting it precisely.

One Cause or Two?

Identifiability is the condition under which something internal — a parameter, mechanism, cause, or hidden state — can in principle be recovered from the observable signal a system makes available. The core claim is uniqueness: among all the internal models you'd allow, the mapping from internals to observations is one-to-one within whatever the observation can actually see. If two distinct internal setups produce the very same observable distribution, the thing is unidentified, and no amount of additional same-kind data can separate them; only a structural change — a new measurement channel, an experiment, a parametric restriction, a prior — can. It is crucial to keep this apart from estimation: identifiability is whether the destination even exists, while estimation noise is the difficulty of getting there once it does. More data, fancier estimators, and bigger computers help only when the object is identifiable in the first place.

 

Identifiability is the structural condition under which an internal unknown — a parameter, mechanism, causal effect, hidden state, or latent variable — is in principle recoverable from the observable signal the system makes available. The defining commitment is a uniqueness claim: across the space of admissible internal models, the mapping from internals to observations is one-to-one within whatever subspace the observation can see. When two distinct internal configurations produce the same observable distribution, the object is unidentified, and no amount of additional data of the same kind can distinguish them — only a structural intervention (a new measurement channel, an experimental manipulation, a parametric restriction, a prior commitment) can. The shape is the same across substrates: a target object you want to know, an observation map the world makes available, and an equivalence class on the target space induced by that map; identifiability is the property that this equivalence class is a singleton for the value of interest. The diagnostic question — could two structurally distinct internals produce identical observations? — is substrate-neutral, and a 'yes' tells you, before any data arrive, that the model is misspecified for the question being asked. What the prime provides is a separation of an information-theoretic upper bound from estimation difficulty: failure to identify is not solved by more data, better estimators, or more computation; estimation noise lives downstream, while identifiability gates whether the destination exists at all.

Structural Signature

the internal target unknown — the observation map the world makes available — the equivalence class the map induces on the target space — the uniqueness (singleton) condition — the information-bound separated from estimation difficulty — the structural levers of enrichment, restriction, and intervention

Identifiability is present (or its failure diagnosable) when each of the following holds:

  • A target unknown (the internal object). A parameter, mechanism, causal effect, hidden state, or latent variable one wants to recover — the thing not directly observed.
  • An observation map (the channel). The mapping from internal configurations to observable distributions — the channel the world makes available.
  • An induced equivalence class (the relating structure). The set of internal configurations that produce the same observable distribution; identifiability is the property that this class is a singleton at the target value of interest.
  • A uniqueness condition (the core invariant). The map is one-to-one within the subspace the observation can see; when two distinct internals yield identical observations the target is unidentified, and the class itself characterizes the irreducible ambiguity.
  • The information-bound separation (the gating invariant). Identifiability is an upper bound on what any data of this kind can reveal, distinct from estimation difficulty downstream; non-identifiability is preserved under more data, bigger models, and more computation.
  • The structural levers (the resolution invariant). A non-singleton class can be collapsed only structurally — enriching the channel (a new measurement), restricting by prior (ruling out members on declared grounds), or intervening (cutting edges to produce a richer observation) — never by more of the same data.

The components compose into a cheap up-front check on the model-and-channel pair, whose dual is anonymity: the identical test with the success criterion flipped to demand a large class.

What It Is Not

  • Not observability. observability asks whether a system's current internal state can be reconstructed from output; identifiability asks whether the model parameters/mechanism can be recovered from input-output data. They are duals in control theory — state-from-output versus model-from-behavior — but distinct objects.
  • Not falsifiability. falsifiability (the nearest neighbor) asks whether a hypothesis could be refuted by some observation; identifiability asks whether a unique internal value can be recovered. A theory can be falsifiable yet have unidentified parameters, or identified yet practically unfalsifiable.
  • Not estimation difficulty. Identifiability is an information-theoretic upper bound on what any data of a kind can reveal; estimation is how hard recovery is given that it is possible. Non-identifiability survives more data, bigger models, and more computation — estimation noise does not.
  • Not confounding. confounding is one cause of non-identifiability of a causal effect (an open back-door path); identifiability is the broader property of unique recoverability, which can fail for many reasons (parameter redundancy, weak instruments) besides confounding.
  • Not identification in the sense of identification/naming. Recovering which member of a set a unique value is (the statistical sense) differs from assigning a label or recognizing an entity (the identification candidate prime). This prime is the recoverability-of-an-internal-unknown sense.
  • Not validity. A measure can be construct-valid yet its parameter unidentified, or identified yet measuring the wrong construct. Identifiability concerns unique recoverability; construct_validity concerns whether the recovered quantity is the intended one.
  • Common misclassification. Blaming sample size — pouring more data, bigger models, or more compute at a non-identified target. Catch it by asking whether two distinct internal configurations could produce identical observation distributions: if yes, the obstacle is structural and upstream of any estimator, and scaling the data is wasted until the channel changes.

Broad Use

Identifiability surfaces wherever something internal must be inferred from something external. In statistics and econometrics it appears as parameter identifiability in structural-equation models, the rank-and-order conditions for simultaneous-equation systems, identification of causal effects via instrumental variables or regression discontinuity, and weak-identification problems where parameters are technically identified but practically not. In causal inference it is the do-calculus identification problem — given a causal DAG and a set of observed variables, when can an interventional distribution be expressed in terms of observed distributions? — and here it is decidable, with an algorithm returning either a formula or a proof of non-identifiability. In control theory and system identification it is the structural and practical identifiability of dynamic-system parameters from input-output data, sitting beside the closely related observability prime. In pharmacokinetics, systems biology, and ecology, many compartmental models are structurally non-identifiable, forcing model reduction or additional measurement channels. In cryptography and security, one-way functions and indistinguishability notions are the deliberate engineering of non-identifiability — between plaintexts given ciphertext, between random and pseudorandom — the same structural pattern with the value-sign flipped: cryptography wants the equivalence class large, science wants it a singleton. In philosophy of mind, the underdetermination of mental content by behavioral output is the identifiability problem applied to the agent's mind. In machine learning, identifiability of factors in nonlinear ICA and of representations under disentanglement is exactly this prime at scale. And in forensics, recovering a generating mechanism from its traces is identifiability of the mechanism from the trace distribution.

Clarity

Naming identifiability separates two questions that working scientists and engineers routinely conflate: can the answer in principle be recovered from this kind of data? and given that it can be recovered, how well can we estimate it with the data we have? The first is an algebraic and structural property of the model-data pair; the second is statistical and numerical. Confusing them produces predictable pathologies — optimizers that converge to plateaus because the loss surface has a non-identified ridge, Bayesian posteriors that equal their priors on a parameter the likelihood cannot move, practitioners blaming sample size for what is actually a degenerate information channel. The prime also clarifies a common false hope: that with enough flexibility ("a bigger model") or enough data the answer will come out. Non-identifiability is preserved under both. If two configurations produce identical observation distributions, neither a larger sample nor a richer model class will tell them apart — only a structural move will. The clarifying force is to route a stuck inference to the correct diagnosis, distinguishing a channel that cannot reveal the target from an estimator that has not yet been pushed hard enough.

Manages Complexity

Identifiability collapses the question "can we learn X from these data?" into a tractable structural-algebraic check, performed on the model-and-channel pair before any computation on the data itself. Methods exist substrate by substrate — rank conditions for structural-equation models, the identification algorithm for causal DAGs, observability and identifiability Gramians for linear dynamic systems, Lie-algebraic tests for nonlinear systems, parameter-redundancy theorems for compartmental models — but they all implement the same structural question on different substrates: is the target-to-observation map injective at the value of interest? The compression is high because the answer is binary in the structural case, or quantifiable as a degree in the practical case, and it licenses up-front decisions: if non-identified, redesign the experiment or restrict the model class; if weakly identified, expect huge confidence intervals and plan accordingly; if strongly identified, proceed to estimation. Identifiability analysis is thus the cheap structural filter that precedes the expensive estimation loop, and recognizing it as a single recurring check across all these substrate-specific tests is itself a major economy — one concept rather than a dozen unrelated technicalities.

Abstract Reasoning

The pattern licenses several substrate-independent moves. Equivalence-class enumeration: given a target and a channel, derive the equivalence class on the target induced by the channel; if it is non-singleton, the target is unidentified and the class itself characterizes the irreducible ambiguity, whether its members are parameter values, Markov-equivalent causal graphs, mental states, or suspect identities. Channel enrichment: non-identifiability is fixed by enlarging the observation map — a new instrument, measurement, or experimental arm — until it distinguishes the previously equivalent configurations, and the cost of enrichment can be weighed against the value of resolved ambiguity to yield an experimental-design criterion. Restriction-by-prior: alternatively the class is collapsed by ruling out members on prior grounds, with the honest move being to declare which members are ruled out and on what grounds, since many non-identifiability disasters are committed silently by a default prior nobody recognized as load-bearing. Intervention as identification: when no passive observation suffices, an active intervention cuts edges in the underlying mechanism and produces a strictly richer observation — the formalization of why experiments identify what observational studies cannot. And the dual of identifiability is anonymity: the same test with the success criterion flipped, so that differential privacy, k-anonymity, and indistinguishability obfuscation specify minimum sizes of the equivalence class rather than demanding a singleton.

Knowledge Transfer

Because identifiability is a bare relational structure — a target, a channel, and the equivalence class the channel induces — a technique developed in one field transfers to any other by re-identifying those three roles, and the prime's reach is exactly the reach of that re-identification. The causal-graph identification algorithm transfers to systems-biology rate-constant recovery: ask which interventions on which pools collapse the parametric equivalence class to a singleton, and the same algorithmic structure that decides do-calculus identifiability now drives experimental-design choices in pharmacology and policy evaluation. Cryptographic indistinguishability transfers to privacy guarantees: the decision-theoretic framing — an adversary cannot beat chance at distinguishing two hypotheses given the observation — carries from cryptographic security to data privacy and A/B-test protection, and even to philosophical thought experiments, because in every case the object is the size of an equivalence class under an observation channel. The econometric notion of weak identification transfers to ML representation learning: the recognition that even technically identified parameters can be practically unidentified by flat likelihood surfaces maps onto flat-minima problems and the practical non-identifiability of disentangled representations, and the corrective move is the same structural one in both — add measurement channels (auxiliary tasks, contrastive pairs) rather than scale up. And the observability/identifiability duality of control theory transfers to model-based experimental design generally: recognizing that state-from-output recovery and model-from-input-output recovery are dual problems with a shared algebraic framework lets a designer jointly shape the input and the measurement channel to maximize the rank of the recoverability operator. In every transfer the practitioner runs the identical diagnosis — name the target, name the channel, derive the induced equivalence class, and choose among enrichment, restriction, and intervention to make it a singleton — and the transfer is secure precisely because none of these steps mentions the substrate: a pharmacologist disambiguating rate constants and a cryptographer guaranteeing indistinguishability are running the same structural test, distinguished only by whether they want the equivalence class small or large.

Examples

Formal/abstract

A two-compartment pharmacokinetic model is the cleanest formal instance, exhibiting non-identifiability and its structural cure. The target unknown is the set of rate constants governing how a drug moves between blood and tissue and is eliminated — say \(\{k_{12}, k_{21}, k_{10}\}\). The observation map is the only thing measurable: the drug concentration in blood over time after a dose. Solving the model shows the concentration curve is a sum of two decaying exponentials, \(C(t) = A e^{-\alpha t} + B e^{-\beta t}\), whose four observable parameters \(\{A, B, \alpha, \beta\}\) are fewer than, or related ambiguously to, the underlying rate constants. The induced equivalence class is therefore non-singleton: distinct triples of rate constants produce identical blood-concentration curves, so the underlying mechanism is structurally unidentified. The prime's gating invariant is decisive and counterintuitive — sampling the blood more densely or for longer (more data of the same kind) cannot separate the indistinguishable triples, because they generate the very same curve; the ambiguity is an upper bound on what this channel can reveal, not an estimation deficiency. The structural levers are exactly what pharmacologists reach for: enrich the channel by adding a tissue-biopsy or urinary-excretion measurement that the competing triples predict differently, restrict by prior (fix one rate constant from independent physiology, declaring that assumption openly), or intervene with a different dosing route that exposes a transfer the original could not. The cheap up-front check — derive the equivalence class on the rate constants before collecting any patient data — saves a trial that would otherwise converge to a flat likelihood ridge and blame sample size.

Mapped back: The compartmental model instantiates every component — target (rate constants), channel (blood concentration), a non-singleton equivalence class, the information-bound that no extra same-kind data can beat, and resolution by enrichment/restriction/intervention — and demonstrates the prime's core separation: identifiability gates whether the answer exists before any estimator is run.

Applied/industry

Differential privacy shows the prime's dual — deliberately engineered non-identifiability — in a data-release substrate. Here the value-sign is flipped: instead of a singleton, the designer wants the equivalence class large. The target unknown is whether a specific individual's record is present in (or contributed to) a released statistic; the observation map is the published query answer (a count, a model, an average). The privacy mechanism adds calibrated random noise so that the induced equivalence class — the set of databases that could have produced the observed output — is large and, crucially, contains both the database with the individual and the one without. The formal guarantee is exactly the prime's uniqueness condition negated: for the adversary, the two neighboring databases are nearly indistinguishable (their output distributions differ by at most a factor \(e^\varepsilon\)), so no analysis of the released data can identify the individual's membership better than a bounded margin over chance. The prime's gating invariant protects the guarantee precisely as in the scientific case but inverted: because indistinguishability is a structural property of the channel, the adversary cannot defeat it with more computation or auxiliary data of the same kind — only a structural change (a different, less private mechanism; a side channel) would shrink the class. The design lever is the privacy budget \(\varepsilon\), which directly sets the minimum size of the equivalence class, trading utility against the guaranteed ambiguity. The same dual structure governs cryptographic indistinguishability (plaintexts kept in one large class given ciphertext) and k-anonymity (each record made indistinguishable among at least \(k\)).

Mapped back: Differential privacy runs the prime as its dual end-to-end — target (membership), channel (noisy release), an equivalence class deliberately enlarged, an information bound the adversary cannot beat with more same-kind data, and \(\varepsilon\) as the lever sizing the class — and shows the prime's reach: the identical target/channel/equivalence-class diagnosis serves the scientist who wants the class a singleton and the privacy engineer who wants it large.

Structural Tensions

T1 — Identifiability versus Estimation Difficulty (Information Bound versus Noise). The prime's central tension is the boundary with estimation: identifiability is an upper bound on what any data of this kind can reveal, distinct from how hard the recovery is given that it is possible. The failure mode is blaming sample size: pouring more data, bigger models, or more computation at a non-identified target, when no amount of same-kind data can separate the equivalent configurations. Diagnostic: ask whether two distinct internals could produce identical observation distributions; if yes, the obstacle is structural and lives upstream of any estimator, so scaling the data is wasted effort until the channel is changed.

T2 — Identification versus Anonymity (Sign Flip of the Same Test). The identical structural test serves opposite goals — science wants the equivalence class a singleton, privacy and cryptography want it large — so the "success" criterion is value-laden and substrate-set. The failure mode is goal-criterion confusion: importing an identification technique into an anonymity setting (or vice versa) while keeping the wrong success direction, so a method built to shrink the class is used where it should be enlarged. Diagnostic: ask whether the application wants the target recovered or hidden; the same target/channel/equivalence-class machinery applies, but the desired class size flips, and a method must be matched to the sign of the goal.

T3 — Channel Enrichment versus Restriction-by-Prior (Resolution Locus). A non-singleton class can be collapsed by enriching the observation channel (a new measurement) or by ruling out members on prior grounds — and the two differ sharply in honesty and cost. The failure mode is silent prior smuggling: collapsing the class with a default prior or modeling assumption nobody recognized as load-bearing, so apparent identification is an artifact of an undeclared restriction. Diagnostic: ask whether uniqueness came from new data or from an assumption; if a parameter is "identified" only because a prior ruled out the competing members, that prior is doing the work and must be declared and defended, not hidden in a default.

T4 — Structural versus Practical Identifiability (Degree versus Binary). Structural identifiability is binary (the map is injective or not), but practical identifiability is a matter of degree — a technically identified parameter can be unrecoverable from a flat likelihood ridge. The tension is between the clean algebraic answer and the noisy finite-sample reality. The failure mode is false comfort from structural identifiability: declaring a target recoverable because the map is injective in principle, while the likelihood surface is so flat that estimates are meaningless. Diagnostic: ask not only whether the class is a singleton but how sharply the observation distinguishes nearby targets; weak identification produces huge confidence intervals and optimizer plateaus that structural analysis alone does not catch.

T5 — Passive Observation versus Active Intervention (Channel Power). Some targets are non-identifiable from any passive observation yet become identifiable under intervention, because cutting edges produces a strictly richer channel. The tension is between the data the world offers and the data an experiment could force. The failure mode is observational resignation: concluding a causal target is unrecoverable when it is merely unidentified from passive data, missing that an intervention would collapse the class. Diagnostic: ask whether an active manipulation (a new arm, a forced value, a different dosing route) would distinguish the equivalent configurations; if the obstacle is the passivity of the channel, the resolution is to intervene, not to abandon the target.

T6 — Target-of-Interest versus Full-Model Identifiability (Scopal Focus). Identifiability is a property of a specific target value, not the whole model — a model can be globally non-identifiable while the particular effect one cares about is identified, or vice versa. The failure mode is scope mismatch: declaring the whole model unidentified and giving up, when the estimand of interest is recoverable, or assuming a needed quantity is identified because the model "mostly" is. Diagnostic: derive the equivalence class on the target of interest, not the entire parameter space; identifiability must be checked at the value the question actually depends on, since a non-identified nuisance parameter need not block an identified effect, and an identified model need not identify every estimand.

Structural–Framed Character

Identifiability sits at the pure structural end of the structural–framed spectrum, with a frontmatter aggregate of 0.0 — every diagnostic reads zero. It is a bare relational structure: a target unknown, an observation map, and the equivalence class the map induces on the target space, with identifiability the property that this class is a singleton at the value of interest. The structural-abstraction subscore of 5/5 reflects the same fact from the other direction — a one-to-one observation map is about as abstract as a relational signature gets.

The pattern is recognized across substrates without importing any framing context or normative load. It carries no home vocabulary that must travel (vocab_travels 0.0): the target/channel/equivalence-class triad is re-identified, not translated, across structural-equation models, causal DAGs, control-theoretic system identification, compartmental pharmacokinetics, cryptographic indistinguishability, and philosophy of mind, each in its own field's words. It carries no evaluative weight (evaluative_weight 0.0): the same test serves the scientist who wants the class a singleton and the privacy engineer who wants it large, so "success" is value-laden only by the application's goal, not the pattern. Its origin is formal (institutional_origin 0.0), a piece of inference theory rather than any institution's product. It is not human-practice-bound (human_practice_bound 0.0): whether a molecule's binding site is recoverable from its traces, or rate constants from a blood-concentration curve, is an identifiability question with no human practice required. And invoking it recognizes rather than imports (import_vs_recognize 0.0): to ask whether a target is identifiable is to inspect a map already there for injectivity, adding no interpretive frame.

The cryptographic dual — deliberately engineered non-identifiability, where the value-sign is merely flipped — is itself evidence of the structural read: the same target/channel/class machinery serves opposite goals across utterly different substrates, which only a frame-free relational pattern could do. The 0.0 aggregate is correct.

Substrate Independence

Identifiability is a strongly substrate-independent prime — composite 4 / 5 on the substrate-independence scale, and its structural abstraction reaches the ceiling at 5: the signature — a target unknown, an observation map, and the equivalence class the map induces on the target space, with identifiability the property that the class is a singleton at the value of interest — is a one-to-one observation map, about as abstract and frame-free as a relational signature gets. It travels cleanly (domain breadth 4): parameter and rank conditions in statistics and econometrics, the do-calculus identification problem in causal inference (where it is decidable), structural and practical identifiability in control theory and system identification, structurally non-identifiable compartmental models in pharmacokinetics and systems biology, deliberately engineered non-identifiability in cryptography (the value-sign merely flipped — the class wanted large rather than singleton), the underdetermination of mental content in philosophy of mind, and disentanglement in machine learning. Transfer evidence is concrete (4): the target/channel/equivalence-class triad is re-identified rather than translated across these fields, and the same machinery serving opposite goals — science wanting a singleton, privacy wanting a large class — is itself strong evidence of frame-free transfer. What holds the composite to 4 rather than 5 is a mild lean toward formal and statistical substrates where the rigorous machinery lives; the relation is medium-neutral, but its concrete documented instances cluster in formal-inference settings rather than spreading as evenly as a canonical physical prime.

  • Composite substrate independence — 4 / 5
  • Domain breadth — 4 / 5
  • Structural abstraction — 5 / 5
  • Transfer evidence — 4 / 5

Neighborhood in Abstraction Space

Identifiability sits in a moderately populated region (42nd percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.

Family — Identity, Reference & Placeholders (10 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-06-14

Not to Be Confused With

The nearest neighbor (similarity 0.91) is falsifiability, and the two are easily merged because both concern what observation can do to a hypothesis — yet they ask structurally different questions. Falsifiability asks whether a hypothesis forbids some observation, so that the observation, if it occurred, would refute it; it is a property of a theory's empirical content, a precondition for its being scientific in the Popperian sense. Identifiability asks whether a unique internal value can be recovered from the observable channel — a property of the model-and-channel pair, indifferent to refutation. The two dissociate cleanly. A theory can be eminently falsifiable while its parameters are non-identified: the model as a whole makes risky predictions, yet several distinct parameter triples produce identical observable distributions, so the data could refute the theory but cannot pin down its internals. Conversely a parameter can be sharply identified within a model that, taken whole, is hard to falsify. Confusing them sends the analyst to the wrong remedy: a falsifiability problem is fixed by deriving a forbidden prediction, an identifiability problem by enriching the channel or restricting the class. One concerns whether observation could prove the theory wrong; the other concerns whether observation could tell you the value.

A second genuine confusion is with observability, its control-theoretic dual. Observability asks whether a system's current internal state can be reconstructed from its outputs over time; identifiability asks whether the model's parameters or structure can be recovered from input-output behavior. These are formally dual — one recovers a state given a known model, the other recovers the model given observed behavior — and they share an algebraic framework (rank conditions, Gramians), which is exactly why they are confused. But the objects differ: observability takes the model as given and asks about the state inside it; identifiability asks about the model itself. The distinction is load-bearing in experimental design, where one may need to shape the input to make parameters identifiable and shape the measurement to make states observable, and these are separate ranks to maximize. Treating them as one concept conflates "can I know where the system is now?" with "can I know how the system works?" — questions with different cures.

A third confusion is with confounding. Confounding is a specific cause of non-identifiability for a causal effect: an open back-door path makes the interventional distribution non-recoverable from observational data. But identifiability is far broader. A causal effect can be non-identified with no confounding at all (parameter redundancy in a compartmental model, a weak or absent instrument, a Markov-equivalent graph class), and confounding is just one of the structural reasons the target-to-observation map fails to be injective. Reducing identifiability to confounding leads to the error of checking only for back-door paths and declaring victory, missing the redundancy or weak-instrument failures that also collapse the equivalence class to non-singleton. Confounding is one route to non-identifiability; identifiability is the general property that route is one way of violating.

For a practitioner the distinctions route a stuck inference correctly. Confusing identifiability with falsifiability looks for a refuting prediction when the problem is a non-injective channel. Confusing it with observability conflates recovering the state with recovering the model. Confusing it with confounding checks only back-door paths and misses redundancy and weak identification. The unifying discipline is the prime's diagnosis: name the internal target, name the channel, derive the equivalence class the channel induces, and — distinguishing this structural question from refutability, from state reconstruction, and from any single cause like confounding — choose among enrichment, restriction, and intervention to make the class a singleton.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.